Skip to content
mini-lab
Log in Sign up
Features of prelude-2

Feature #2036

Fires on “?” (not labeled yet: read the examples)

Fires on
0.41% of tokens
Mostly in
python
Layer
after block 4

Where it fires

  • python 46%
  • agent 38%
  • chat 10%
  • worksheet 4%
  • story 2%

Makes the model say

`#⏎ assert$ drink README H

Makes it not say

ings boughtningving Now tookLook he

Strongest examples

On text the dictionary never trained on. The underlined token is where it fires most.

  1. agent 2.94
    ?<|user_end|><|assistant_start|><|tool_call_start|>s
  2. chat 2.94
    ?<|assistant_end|><|user_start|>Hello!
  3. worksheet 2.94
    ?⏎171
  4. chat 2.94
    ?<|user_end|><|assistant_start|>Once upon
  5. chat 2.94
    ?<|user_end|><|assistant_start|>I'm
  6. chat 2.94
    ?<|user_end|><|assistant_start|>The answer
  7. chat 2.94
    ?<|user_end|><|assistant_start|><|tool_call_start|>calculator
  8. chat 2.94
    ?<|user_end|><|assistant_start|>The answer
  9. chat 2.94
    ?<|user_end|><|assistant_start|><|think_start|>1
  10. chat 2.94
    ?<|user_end|><|assistant_start|><|tool_call_start|>calculator
  11. chat 2.94
    ?<|user_end|><|assistant_start|>Once upon
  12. chat 2.43
    today?<|assistant_end|><|bos|># opencode.